One of the fundamental tasks in information theory is the compression of information. To achieve this in the quantum domain, quantum autoencoders that aim to compress quantum states to low-dimensional ones have been proposed. When taking a pure state as the reference state, there exists an upper bound for the encoding fidelity. This bound limits the compression rate for high-rank states that have high entropy. To overcome the entropy inconsistency between the initial states and the reconstructed states, we allow the reference state to be a mixed state. A new cost function that combines the encoding fidelity and the quantum mutual information is proposed for compressing general input states. In particular, we consider the reference states to be a mixture of maximally mixed states and pure states. To achieve efficient compression for different states, two strategies for setting the ratio of mixedness (in the mixture of maximally mixed states and pure states) are provided based on prior knowledge about quantum states or observations obtained from the training process. Numerical results on thermal states of the transverse-field Ising model, Werner states, and maximally mixed states blended with pure states illustrate the effectiveness of the proposed method. In addition, quantum autoencoders using mixed reference states are experimentally implemented on IBM Quantum devices to compress and reconstruct thermal states and Werner states.